Abstract

Many modeling studies have indicated that deforestation will increase the average annual temperature in the Amazon. However, few studies have investigated the potential for deforestation to change the frequency and intensity of extreme events. This problem is addressed here using a variable-resolution GCM. The characteristic length scale (CLS) of the model’s grid mesh over South America is 25 km, comparable to that used by limited-area models. For computational efficiency, the CLS increases to 200 km over the rest of the world. It is found that deforestation induces large changes in the frequency of wintertime extreme cold events. Large increases in cold event frequency and intensity occur in the western Amazon and, surprisingly, in parts of southern South America, far from the actual deforested area. One possible mechanism for these remote effects involves changes in the position of the subtropical jet, caused by temperature changes in the Amazon. Increased understanding of these potential changes in extreme events will be important for local agriculture, natural ecosystems, and the human population.

1. Introduction

Many modeling studies have indicated that the large-scale deforestation of the Amazon would have substantial climatic implications (e.g., Henderson-Sellers et al. 1993; Lean and Rowntree 1993; Gash and Nobre 1997; Hahmann and Dickinson 1997; Costa and Foley 2000; Gedney and Valdes 2000; Werth and Avissar 2002; Findell et al. 2006; Sampaio et al. 2007; Hasler et al. 2009; Medvigy et al. 2011). Decreases in average annual precipitation and increases in average annual surface temperature in the Amazon region have been two of the most commonly reported results, implying changes in the probability density functions (PDFs) of surface temperature and rainfall. However, deforestation may cause PDFs to change in ways that are more complex than simple shifts in the mean. For example, a large increase in temperature variance may cause an increase in extreme cold events even if mean temperature increases.

GCMs have rarely been used to investigate the impact of deforestation on extreme events because their typical horizontal grid cell size (2°–10°) smoothes out a considerable amount of spatial variability, leading to underestimates of extreme event frequency (Diffenbaugh et al. 2005; Kunkel et al. 2010). To capture smaller scales within the context of a GCM simulation, Medvigy et al. (2011) used an atmospheric GCM, the ocean–land–atmosphere model (OLAM; Walko and Avissar 2008a,b, 2011; Medvigy et al. 2008, 2010, 2011), capable of simulating different length scales with either a triangular or hexagonal numerical grid footprint (Walko and Avissar 2011). They simulated most of South America using a triangular grid mesh with a 25-km characteristic length scale while simulating much of the rest of the world with a grid mesh having a more computationally efficient characteristic length scale of 200 km. For South America, they argued that this model configuration yielded precipitation PDFs that compared favorably to those from the Tropical Rainfall Measuring Mission (Kummerow et al. 1998). While Medvigy et al. (2011) focused on changes in precipitation mean, intensity, and frequency, they also suggested that the frequency of extreme cold events in South America could increase following deforestation.

Although cold air incursions into southern South America are observed to occur year-round, they are stronger and more common during the austral winter, June–August (JJA) (Marengo et al. 1997a). Garreaud (2000) described the typical structure of a cold event, which we now summarize. Prior to a strong cold event, it is typical for an anticyclonic high pressure center to move eastward from the Pacific Ocean to southern South America, where it is blocked by the Andes. It then propagates southward to about 45°S, where the altitude of the Andes becomes sufficiently low for it to cross into southern South America. Over the course of several days, the anticyclone is channeled northward through Argentina along the eastern slopes of the Andes to about 25°S. At this point, the anticyclone attains an eastward velocity, sweeps across Brazil, and moves toward the Atlantic. Individual events adhere to this paradigm to varying degrees; details on the spatial extent and persistence of individual events have previously been described (Marengo et al. 1997a; Garreaud 1999, 2000; Müller et al. 2005; Müller and Berri 2007). Overall, cold air incursions can occur up to several times a year. They have direct implications for the coffee crop of southern Brazil (Marengo et al. 1997a), and occasional frosts in Argentina and southern Brazil (Marengo et al. 1997a,b; Garreaud 2000; Müller et al. 2005; Müller 2010) can also impact harvests of wheat, citrus, and soybeans.

The objective of this paper is to use a numerical model to determine if large-scale deforestation can change the sensitivity of South America to extreme cold events during June–August. It builds on Medvigy et al. (2011) by providing more robust estimates of the statistical significance of impacts and by analyzing changes in the synoptic evolution of extreme events. In section 2, we describe the numerical model, simulation design, and statistical approach. In section 3, we discuss the model’s characterization of extreme cold events when land cover from the early 1990s is used as a boundary condition. Results from the National Centers for Environmental Prediction (NCEP) reanalysis are presented for comparison. In section 4, we describe the simulated changes in extreme events that result from deforestation. Statistical significance and changes in the synoptic evolution of extreme events are both evaluated. Our conclusions are presented in section 5.

2. Methods

a. Model simulations

We use the OLAM (Walko and Avissar 2008a,b, 2011) run as an atmospheric GCM with prescribed sea surface temperatures. An advantage of OLAM is that its grid mesh may employ multiple characteristic length scales. We simulate most of South America at 25-km characteristic length scale, and then allow the grid mesh to gradually coarsen to about 200-km characteristic length scale far from South America. This enables us to capture regional-scale circulations without the need for lateral boundary conditions, while maintaining a tractable computational cost (Medvigy et al. 2008, 2010, 2011). The vertical grid consists of 53 levels, stretching from 200 m near the surface to 2 km near the model top at 45 km. Amazon deforestation has previously been simulated with OLAM (Medvigy et al. 2008, 2011), and the model’s ability to simulate South America’s hydroclimate has previously been evaluated (Medvigy et al. 2008, 2010, 2011). The model’s land surface parameterization (Walko et al. 2000) and other parameterizations are the same as those used in Medvigy et al. (2011).

We consider two land cover scenarios, both taken from Medvigy et al. (2011). In our control run (“CON”), depicted in Fig. 1a, each land grid cell is assigned a single land cover classification according to the Olson Global Ecosystem framework (Olson 1994a,b), which was based on satellite imagery from 1992 to 1993. About 10% of the Amazon is classified as agriculture or short grass in CON. A second simulation (“TOT”), depicted in Fig. 1b, is identical to CON in every way except land cover classification. TOT, meant to represent the total deforestation of the Amazon, classifies all land grid cells between 75°–49°W and 15°S–0° as deforested. The land surface and vegetation properties of these deforested grid cells are prescribed according to in situ measurements at pasture sites (Gash and Nobre 1997) and have been tested in previous studies (Gandu et al. 2004; Avissar and Werth 2005; Ramos da Silva et al. 2008; Hasler et al. 2009; Medvigy et al. 2011). A more sophisticated treatment might distinguish between pasture, soy, and cultivation of other crops, but we expect differences between these types to be much smaller than the differences between tropical forest and pasture (Sampaio et al. 2007).

Fig. 1.

Maps of land cover for the (a) CON and (b) TOT simulations. Color bar correspondence is as follows: 1 = deciduous shrub, 2 = deforested, 3 = evergreen broadleaf forest, 4 = water, 5 = short grass, 6 = crops and mixed farming, 7 = wooded grassland, and 8 = other.

Fig. 1.

Maps of land cover for the (a) CON and (b) TOT simulations. Color bar correspondence is as follows: 1 = deciduous shrub, 2 = deforested, 3 = evergreen broadleaf forest, 4 = water, 5 = short grass, 6 = crops and mixed farming, 7 = wooded grassland, and 8 = other.

Atmospheric and soil initial conditions are prescribed from NCEP reanalysis from 0000 UTC 1 October 1996 (Kalnay et al. 1996). The CON and TOT simulations are forced with weekly, 1° sea surface temperatures (SSTs) (Reynolds et al. 2002), and sea ice extent from NCEP reanalysis (Kalnay et al. 1996). CO2 and other greenhouse gas concentrations are held fixed throughout the simulations at current-day levels. However, several recent studies of deforestation using atmospheric GCMs have been forced with climatological SSTs (Costa et al. 2007; Sampaio et al. 2007; Hasler et al. 2009). This approach may not be appropriate for the current work because extreme cold events in South America have been linked to SST anomalies (Müller 2010). To test this idea, we construct an average annual cycle of SSTs using the 1996–2010 observations. We then carry out two additional simulations, “CONAVG” and “TOTAVG,” which are identical to CON and TOT, respectively, except that they are forced with the average annual cycle of SSTs in each year. We simulate the period from 1 October 1996 to 31 December 2010. Simulations using climatological SSTs indicate that soil moisture and soil temperature in the Amazon region equilibrate after about a year, and so all days in 1996 and 1997 are discarded as spinup, leaving 13 yr for analysis. For comparison, Medvigy et al. (2011) analyzed 8 yr.

b. Definition and detection of extreme events

Bell et al. (2004) used specific threshold temperatures to define extreme cold events. Here, we use their “T05” to quantify extreme cold events. We define this metric on a seasonal basis. In each JJA period, we construct the cumulative distribution function of minimum daily temperature, and identify the temperature corresponding to the fifth percentile. For an N-yr period, there are N such temperatures. We average these N temperatures to obtain T05.

We perform a series of statistical tests to evaluate the significance of differences in CON and TOT extreme event frequency. All statistical tests are conducted in R (R Development Core Team 2008). Comparing JJA T05 from CON and TOT, we have 2 × 13 samples for each model grid cell (in each of the 13 simulated years, there is one sample from CON and one sample from TOT). A t test can be used to test the null hypothesis that the means of the CON and TOT samples are equal, provided that the CON (and TOT) samples are independent and normally distributed. If the normality assumption does not hold, a nonparametric test such as the Wilcoxon signed-rank test (“wilcox.test” in R; Hollander and Wolfe 1999) may be used instead of the t test. The null hypothesis of the Wilcoxon signed-rank test is that the median difference between the CON and TOT samples is zero, and it assumes that the “TOT minus CON” differences are independent and identically distributed. We used the Shapiro–Wilk test (“shapiro.test” in R; Royston 1982) to test for normality and the Ljung–Box test (“Box.test” in R; Ljung and Box 1978) to test for independence. The 95% confidence level is taken as the threshold for statistical significance throughout this paper.

In simulations with interannually varying SSTs, extreme cold event frequency may depend on a particular year’s SST forcing, and “TOT minus CON” differences may not be identically distributed. With our relatively small sample size, statistical tests of whether samples are identically distributed have little power. However, it is possible to test the null hypothesis that the “TOT minus CON” differences are stationary using a Kwiatkowski–Phillips–Schmidt–Shin (KPSS) test (“kpss.test” in R; Kwiatkowski et al. 1992). Not rejecting this hypothesis is a necessary, but not sufficient, condition for the samples to be identically distributed. A way to address this more rigorously would be to generate an ensemble of simulations, forced with the same SST history but with slightly different initial conditions. However, this approach is beyond our current computational abilities, and may not even be necessary if extreme event frequency is primarily sensitive to the high-frequency SST variability that occurs in every year (Lloyd and Vecchi 2010; Barron et al. 2011, manuscript submitted to Geophys. Res. Lett.) rather than multiyear trends in SSTs.

3. Characterization of South American cold events: Control simulation

a. Model comparisons

We focus here on characterizing JJA T05 and T05 event frequency in South America. Although comparisons of the CON simulation with weather station data can be used to build confidence in the OLAM’s ability to simulate extreme events, such comparisons are very complex for South America because of the sparse distribution of weather stations and inhomogeneities in long-term records. One existing dataset that has been designed to carefully control for these issues is the HadGHCND gridded temperature dataset (Caesar et al. 2006). It is based upon near-surface maximum and minimum temperature observations and is intended for the analysis of climate extremes and also for climate model evaluation. However, no South American grid cell in this dataset includes more than 18% of the daily minimum temperature values for JJA days in the decade 1998–2007, and no grid cell in the Amazon has any data. This sampling is too limited to allow unbiased calculations of T05 for South America, and thus reflects the strong need for the collection of additional weather station data in South America.

Because observations are so limited, we have compared the CON simulation with NCEP reanalysis. The reanalysis has complete coverage of South America with no missing values, facilitating the computation of T05 statistics. The reanalysis T05 statistics are nevertheless far from perfect because of the dearth of South American surface or upper-air data available for assimilation and because the horizontal resolution (2.5°) is relatively coarse. Because of the coarse resolution, mesoscale circulations and topography are both poorly represented in the reanalysis. Thus, comparison of the CON simulation with reanalysis is somewhat crude, but it is still worthwhile because OLAM’s temperature climatology has not been previously evaluated.

Before examining T05, we first consider daily average temperature. This metric is important because daily average temperature from CON has not previously been evaluated, and also because daily average temperatures provide a context for extremes. We computed the average temperature for all days in JJA in the years 1998–2010 (for a total of 1196 days). For brevity, we subsequently refer to this as a “JJA average.” Over land, the JJA average temperature from CON (Fig. 2a) takes its maximum values along the northern coast of South America and in the eastern Amazon. The Andes are well-resolved, and are characterized by temperatures <5°C. Locally cold areas are also apparent near the Guyana Highlands and the Brazilian Highlands. Southernmost Argentina and Chile experience near-freezing temperatures. For comparison, temperatures derived from the reanalysis are shown in Fig. 2b. The coarse spatial resolution is obvious here; nevertheless, the reanalysis temperatures generally fall within 2°C of the CON temperatures for much of South America. The most prominent exception is the eastern flank of the Andes, where the reanalysis gives temperatures 3°–8°C cooler than the CON simulation. Along the Andes themselves, the CON simulation is about 5°C cooler than the reanalysis. Given the difference in spatial resolution in CON and the reanalysis, these differences near the Andes are unsurprising. CON is also somewhat cooler than reanalysis over the oceans. This may be because the temperature in CON is representative of a grid box with 200-m vertical extent, and the reanalysis is a 2-m temperature product. Because the marine boundary layer top is often near the 200-m level or lower, it is relatively easy for cold air aloft to mix into the lowest model grid box over the oceans.

Fig. 2.

Daily average temperature (°C) averaged over all days in June–August for the years including 1998–2010. (a) The OLAM CON simulation and (b) NCEP reanalysis are shown.

Fig. 2.

Daily average temperature (°C) averaged over all days in June–August for the years including 1998–2010. (a) The OLAM CON simulation and (b) NCEP reanalysis are shown.

Moving on to extreme event statistics, Fig. 3 presents the JJA-averaged T05 from CON (Fig. 3a) and the reanalysis (Fig. 3b). In both cases, the spatial patterns of T05 are similar to the spatial patterns of average daily temperature (cf. Fig. 2). The largest differences between CON and the reanalysis are again located along the eastern flank of the Andes, where the reanalysis is 8°–12°C cooler than CON. Along the peaks of the Andes, the T05 from CON is about 10°C cooler than the T05 from the reanalysis. Otherwise, in northern South America, CON and the reanalysis T05 differ by <2°C. In northeastern Argentina and southern Brazil, the reanalysis T05 is 2°–3°C cooler than the T05 from CON.

Fig. 3.

June–August T05 statistics are shown with temperatures in °C. (a) T05 from the CON simulation, (b) T05 from the reanalysis, (c) average T05 event frequency minus five for CON, (d) average T05 event frequency minus five for the reanalysis, (e) standard deviation of T05 event frequency for CON, and (f) standard deviation of T05 event frequency for the reanalysis.

Fig. 3.

June–August T05 statistics are shown with temperatures in °C. (a) T05 from the CON simulation, (b) T05 from the reanalysis, (c) average T05 event frequency minus five for CON, (d) average T05 event frequency minus five for the reanalysis, (e) standard deviation of T05 event frequency for CON, and (f) standard deviation of T05 event frequency for the reanalysis.

We also computed the average and standard deviation of T05 event frequency for the CON simulation and the reanalysis. The average T05 event frequency is close to 5 for both CON (Fig. 3c) and the reanalysis (Fig. 3d), and both estimates of average T05 event frequency exceed 5 in the eastern tropical Pacific. This deviation from 5 occurs because the PDF of T05 event frequency is not symmetric around 5 and is instead skewed toward higher event frequencies. CON (Fig. 3e) and the reanalysis (Fig. 3f) also have similar values for the standard deviation of T05 event frequency. Typical values over South America were 2–4. However, CON has larger values than the reanalysis in southeastern Brazil and smaller values than the reanalysis in the central Amazon. Overall, we are satisfied with the comparison between CON and the reanalysis temperatures and are not surprised to find the largest differences near the Andes.

b. Regional-scale T05 events

We now describe regional-scale extreme cold events in the CON simulation. While it is straightforward to identify such events for an individual grid cell using T05, it is less obvious how to define a regional-scale extreme cold event. We proceed by creating a new binary variable, “COLDFLAG.” COLDFLAG is defined for each grid cell for each day in JJA. Its default value is zero. However, if a grid cell’s daily minimum temperature falls below its JJA T05, COLDFLAG is assigned a value of one for that grid cell for that day. We then perform an empirical orthogonal function (EOF) analysis on COLDFLAG for all grid cells in the region between 40°S–10°N and 90°W–30°W. The intent of this analysis is to identify the spatial structures that carry relatively large fractions of the variance in COLDFLAG. We note that actual extreme cold events may be linear combinations of all EOFs, and that the principal components of the EOFs, while orthogonal, may not be independent. Furthermore, the EOF analysis is sensitive to exactly which grid cells are included. For example, if only land grid cells were used, we would expect each mode to capture a larger fraction of the total variance.

The first EOF is shown in Fig. 4a. This mode explains 13.4% of the total variance. It has large values over much of the continent, and especially along a transect connecting the western Amazon to southeast Brazil. The second EOF (Fig. 4b) explains 8.5% of the total variance, and has opposite signs for eastern Brazil and southern South America. We computed the cross correlation between the corresponding principal components (PCs) 1 and 2. There was a positive correlation (~0.3) between PC2 lagged by 1–2 days and PC1, indicating that the spatial pattern of EOF2 precedes the spatial pattern of EOF1 by 1–2 days. We also found a negative correlation (~−0.35) between PC2 negatively lagged by 2–4 days and PC1. This indicates that the negative of the spatial pattern of EOF2 tends to follow the spatial pattern of EOF1. To the extent that extreme cold events can be captured by these two EOFs, we would expect cold events to 1) enter South America in the south, 2) move northward, and then 3) move eastward. This describes a trajectory similar to what has been observed (Marengo et al. 1997a; Garreaud 2000; Müller et al. 2005; Müller 2010). In reality, extreme cold events are observed to weaken as they proceed eastward; however, this aspect cannot be evaluated on the basis of Fig. 4 because COLDFLAG is a grid cell-relative quantity.

Fig. 4.

EOFs of COLDFLAG: (a) CON simulation EOF1, (b) CON simulation EOF2, (c) TOT simulation EOF1 (d) TOT simulation EOF2, (e) Difference between EOF1 from TOT and EOF1 from CON, and (f) difference between EOF2 from TOT and EOF2 from CON. Note the different scales in (e),(f).

Fig. 4.

EOFs of COLDFLAG: (a) CON simulation EOF1, (b) CON simulation EOF2, (c) TOT simulation EOF1 (d) TOT simulation EOF2, (e) Difference between EOF1 from TOT and EOF1 from CON, and (f) difference between EOF2 from TOT and EOF2 from CON. Note the different scales in (e),(f).

As noted above, extreme cold events may be linear combinations of all the EOFs. However, because of the cross correlation between PC1 and PC2, we expected that the timing of the peaks of PC1 would correlate with the timing of regional-scale cold events. We therefore generate a composite event by computing the cumulative distribution function of PC1, and identifying the value of PC1 at the 95th percentile. The days at the 95th and higher percentiles corresponded to the 24 largest peaks in PC1 (on average, each peak spanned about 2.5 days). Let di denote the day of peak i, and let T(di) denote the average temperature on day i. We then compute the composite cold event temperature (CCT):

 
formula

The index j represents the time evolution of the composite event. For example, the average “peak day” temperature is CCT0; the average temperature two days before a peak is CCT−2; and the average temperature two days after a peak is CCT2. The time evolution of the composite cold event is illustrated in Fig. 5. In the figure, all CCTj are shown relative to the average JJA temperature. CCT−3 and CCT−2 are both several degrees cooler than average for areas south of about 20°S. The cooling intensifies and spreads northward on days CCT−1 and CCT0. These days also have warming in Chile. On days CCT1 and CCT2, the cooling is less intense and is located in the western Amazon and eastern Brazil. Southern South America experiences warming. This behavior is largely consistent with observed extreme cold events (Marengo et al. 1997a; Garreaud 2000; Müller et al. 2005; Müller 2010).

Fig. 5.

Composite cold event temperature, CON simulation. All temperatures are reported in °C relative to the June–August average temperature in each grid cell. (a) CCT−3, (b) CCT−2, and (c) CCT−1 . These are the days immediately preceding the peak in PC1. (d) The day of the peak in PC1 (CCT0) is represented. (e),(f) The days following the peak in PC1 (CCT1 and CCT2, respectively).

Fig. 5.

Composite cold event temperature, CON simulation. All temperatures are reported in °C relative to the June–August average temperature in each grid cell. (a) CCT−3, (b) CCT−2, and (c) CCT−1 . These are the days immediately preceding the peak in PC1. (d) The day of the peak in PC1 (CCT0) is represented. (e),(f) The days following the peak in PC1 (CCT1 and CCT2, respectively).

4. Impacts of deforestation on extreme events

a. Changes in T05

We tested the statistical significance of the changes in wintertime (JJA) T05 event frequency. For each grid cell and year, we computed the difference “TOT minus CON” in T05 event frequency (ΔT05FREQ). For each grid cell, we tested the ΔT05FREQ samples for independence (Box–Ljung test) and stationarity (KPSS test). Null hypotheses of stationarity and independence were rejected for less than 4% of the area in the deforested region (15°S–0°; 75°–49°W), southern South America (South America land between 55°S and 20°S), and northern South America (South American land between 20°S and 12°N). These results are consistent with the assumption of independent and identically distributed samples for most grid cells. The commonly-used t test additionally requires normally distributed samples. We tested the null hypothesis that samples are normally distributed using a Shapiro–Wilk test. The fractional area for which the normality hypothesis was rejected at the 95% confidence level was aggregated for northern South America, southern South America, and the deforested region. These fractional areas were 0.19, 0.24, and 0.17, respectively. Because these areas are larger than what would be expected merely by chance, we will not make any assumptions about normality in the remainder of this work.

An alternative to the t test that does not require normally distributed samples is the nonparametric Wilcoxon test, whose null hypothesis is that the median value of ΔT05FREQ is zero. We performed a paired Wilcoxon test for the TOT and CON T05 samples for each grid cell and found that the null hypothesis was rejected in large areas of South America (Fig. 6a). The western Amazon had a large positive ΔT05FREQ. The increase exceeded 10 days yr−1 in some grids cells, which represented a tripling of extreme cold events. The eastern Amazon had a negative ΔT05FREQ of about 5 days yr−1, indicating the complete absence of extreme cold events. We also found significant changes in locations hundreds of kilometers away from the deforested area. In parts of southern South America, and especially Uruguay and northern Argentina, ΔT05FREQ was about +5 days yr−1, which represented a doubling.

Fig. 6.

(a) Significant changes in June–August T05 event frequency (“TOT minus CON”), with the rectangle indicating the deforested region. (b) Significant changes in June–August T05 temperature (°C; “TOT minus CON”). Gray areas had no significant difference between TOT and CON.

Fig. 6.

(a) Significant changes in June–August T05 event frequency (“TOT minus CON”), with the rectangle indicating the deforested region. (b) Significant changes in June–August T05 temperature (°C; “TOT minus CON”). Gray areas had no significant difference between TOT and CON.

Statistically significant “TOT minus CON” differences in T05 temperature (ΔT05) are shown in Fig. 6b. ΔT05 was on the order of −1°C in the western Amazon, northern Argentina, Uruguay, and southernmost Brazil, but ΔT05 was positive in the eastern Amazon. In grid cells where ΔT05FREQ and ΔT05 were both significant, positive ΔT05FREQ was always associated with negative ΔT05, and negative ΔT05FREQ was always associated with positive ΔT05.

These changes in extreme event statistics in part reflect changes in mean temperature and humidity. The 1500 m (note that OLAM uses a Cartesian vertical coordinate and that this height roughly corresponds to 850 hPa) JJA temperature difference is shown in Fig. 7a. Average temperature is higher in TOT than in CON in the eastern and central Amazon, and lower in TOT than in CON along the eastern flank of the Andes and in southern South America. We also see that the JJA specific humidity was lower in TOT than in CON in most of the Amazon (Fig. 7b). The increase in average temperature and the decrease in average specific humidity in the eastern Amazon are unsurprising; many previous deforestation studies have suggested that deforestation reduces evaporative cooling, leading to increased temperatures and reduced moisture. However, the decreases in average temperature (and corresponding negative values of ΔT05 and positive values of ΔT05FREQ) were unexpected. A possible mechanism for these changes is presented below in section 4c.

Fig. 7.

(a) Temperature difference (°C) and (b) specific humidity difference (g kg−1) between TOT and CON at 1500 m (~850 hPa), averaged over all June–August days in 1998–2010. Gray areas represent topography, and the deforested region is bounded by the rectangle.

Fig. 7.

(a) Temperature difference (°C) and (b) specific humidity difference (g kg−1) between TOT and CON at 1500 m (~850 hPa), averaged over all June–August days in 1998–2010. Gray areas represent topography, and the deforested region is bounded by the rectangle.

b. Changes in the synoptic structure of cold events

We computed the COLDFLAG EOFs of the TOT simulation (Figs. 4c,d). The first and second EOFs explained 14.1% and 8.9% of the total variance, respectively, and were similar to the corresponding percentages from the CON simulation (13.4% and 8.5%, respectively). One notable difference between the CON and the TOT EOFs was that the ridge of high EOF1 values extending from eastern Ecuador to the coast east of Paraguay in CON was shifted to the west in TOT, where it mostly follows the eastern side of the Andes. In addition, more area is covered by positive EOF1 values in CON than in TOT in South America north of the equator and south of 25°S (Fig. 4e). In EOF2, CON and TOT exhibit a qualitatively similar north–south dipole, but EOF2 is less positive in Paraguay and more negative in Colombia in TOT relative to CON (Fig. 4f).

We generated a composite cold event from the TOT simulation using Eq. (1). For easy comparison with the composite cold event from the CON simulation (Fig. 5), we likewise subtracted the average JJA temperature from the CON simulation. The result is shown in Fig. 8. Between Day −4 and Day −2, there are lower temperatures in TOT than in CON for many areas south of about 20°S. By Day −1, TOT has 1°–2°C lower temperatures than CON for central South America. This temperature difference intensifies and expands northward to about 10°S by Day 0. Temperatures continue to be colder in TOT than in CON on Days 1 and 2. However, northeast Brazil is one area where the opposite is true. Overall, these results suggest that the regional-scale cold events identified using EOF1 in the TOT simulation are colder than the regional-scale cold events identified using EOF1 in the CON simulation for most of South America.

Fig. 8.

As in Fig. 5, but for the TOT simulation.

Fig. 8.

As in Fig. 5, but for the TOT simulation.

c. Mechanisms linking deforestation and extreme cold events

In accord with most previous studies, we have found that deforestation in the Amazon causes locally increased temperatures (Fig. 7a). This temperature change implies an increase in the north–south temperature gradient in South America. Intriguingly, this is exactly the opposite effect from that expected to result from increases in greenhouse gas concentrations. In the latter case, a reduced north–south temperature gradient occurs because the poles warm more strongly than the tropics (Meehl et al. 2007). Many previous GCM studies (e.g., Hall et al. 1994; Kushner et al. 2001; Yin 2005) have also found that increases in greenhouse gases result in a poleward shift in the storm tracks and the westerly jet. Here, we propose that Amazon deforestation will have the opposite effect: an increase in the north–south temperature gradient which causes an equatorward shift in the jet. Examination of the JJA-averaged 9200 m (~300 hPa) winds from CON (Fig. 9a) indicates a clear jet entrance region over northern Argentina, Uruguay, and southern Brazil. Figure 9b, which shows the difference in wind speed between TOT and CON, indicates that average jet velocities are ~1 m s−1 larger in TOT than in CON at ~20°S. There is a corresponding reduction in jet speed farther south, around 30°S. These results indicate that the jet has indeed shifted equatorward in response to Amazon deforestation.

Fig. 9.

JJA-averaged winds at 9200 m (~300 hPa). The units are m s−1. (a) Wind vector (arrows) and speed (contours) from the CON simulation and (b) “TOT minus CON” difference in wind speed are shown.

Fig. 9.

JJA-averaged winds at 9200 m (~300 hPa). The units are m s−1. (a) Wind vector (arrows) and speed (contours) from the CON simulation and (b) “TOT minus CON” difference in wind speed are shown.

The importance of upper-level jet entrance regions for subtropical cold surges has already been noted for South America (Garreaud 1999, 2000), Central America (Schultz et al. 1998), and Southeast Asia (Lau and Chang 1987). One mechanism connecting cold events and jets, described by Uccellini and Johnson (1979), proposes that acceleration of the upper-level flow near the jet entrance causes a secondary circulation in the plane normal to the jet axis. Downward motion occurs on the poleward side of the jet, and upward motion occurs on the equatorward side of the jet. Because midtropospheric subsidence tends to increase anticyclonic vorticity at lower levels, this strengthens the surface high pressure center over the continent. Furthermore, rising motions on the equatorward side induce midlevel adiabatic cooling and favor the incursion of cold air northward. Garreaud (1999) used a mesoscale model to simulate a particular South American cold event, and found that this mechanism did indeed play a critical in allowing the cold event to penetrate deeply northward toward the equator.

Because we simulated a stronger jet in TOT than in CON at ~20°S, we expected that a corresponding secondary circulation would be induced near this latitude in TOT. To test this, we computed the JJA-averaged vertical velocity at the 4950 m (~555 hPa) level. Differences between TOT and CON (Fig. 10) are characterized by relative subsidence to the south of the jet and relative upward motion to the north of the jet and are thus in accord with our expectations. We note that this mechanism becomes effective only once the surface anticyclone begins to overlap with the region where the jet is strengthened and not during the early phases of the cold event when the anticyclone is initially crossing over the Andes into South America. This is consistent with Fig. 6, which does not show increases in T05 event frequency or intensity where the anticyclones cross over from the Pacific into the continent.

Fig. 10.

Vertical velocity difference (cm s−1) between TOT and CON at 4950 m (~550 hPa), averaged over all June–August days in 1998–2010.

Fig. 10.

Vertical velocity difference (cm s−1) between TOT and CON at 4950 m (~550 hPa), averaged over all June–August days in 1998–2010.

d. Impacts of sea surface temperatures on T05 events

Previous GCM (Rind et al. 2001) and reanalysis-based (Barros et al. 2002) studies reported that southern South America should be relatively cool when the eastern tropical Pacific is relatively cool. However, these studies did not assess the impact of SSTs on extreme cold events. To determine if there is any spatial pattern of SST anomalies characteristic of extreme cold events, we computed the following sum for each grid cell:

 
formula

SSTobs(di) is the observed SST at the ith peak of PC1 from the CON simulation, and SSTavg(di) is the climatological value of SST at the ith peak of PC1 from the CON simulation. Thus, ΔSSTcon can be thought of as the SST anomaly corresponding to an extreme cold event. We also computed a quantity ΔSSTtot on the basis of the peaks in PC1 of the TOT simulation. Both the ΔSSTcon and the ΔSSTtot anomalies have large negative values in the eastern tropical Pacific (Fig. 11), supporting previous associations of such anomalies with cold temperatures in southern South America (Rind et al. 2001; Barros et al. 2002). ΔSSTcon and ΔSSTtot are also both negative off the coast of southeastern South America and in the Gulf of Alaska, and they are both positive off the coast of the northeastern United States. The only place where the anomalies are markedly different is near Japan.

Fig. 11.

Sea surface temperature anomalies (°C) corresponding to the COLDFLAG PC1 peaks in the (a) CON simulation and (b) the TOT simulation.

Fig. 11.

Sea surface temperature anomalies (°C) corresponding to the COLDFLAG PC1 peaks in the (a) CON simulation and (b) the TOT simulation.

Given this association of extreme cold events with SSTs, we expected that CON and TOT would differ markedly from simulations where each year was driven with the mean annual cycle of SSTs (CONAVG and TOTAVG). We denote the “TOTAVG minus CONAVG” difference in T05 event frequency as ΔT05FREQ-AVG . Like ΔT05FREQ, ΔT05FREQ-AVG was negative in the eastern Amazon (Fig. 12). However, ΔT05FREQ-AVG was not significantly different from zero in the western Amazon. Moreover, ΔT05FREQ-AVG was significantly positive for parts of eastern Brazil. This was not the case in our simulations forced with observed SSTs. In addition, ΔT05FREQ-AVG in southern South America was not as strongly impacted as ΔT05FREQ. These results indicate a strong nonlinear dependence of South American extreme cold events on spatial patterns of SSTs.

Fig. 12.

Significant changes in JJA T05 event frequencies (“TOTAVG minus CONAVG”) for simulations forced with average SSTs from 1996–2010. Gray areas do not have significant changes, and the rectangle indicates the deforested region.

Fig. 12.

Significant changes in JJA T05 event frequencies (“TOTAVG minus CONAVG”) for simulations forced with average SSTs from 1996–2010. Gray areas do not have significant changes, and the rectangle indicates the deforested region.

5. Discussion and conclusions

This study represents an initial attempt to characterize changes in extreme events resulting from deforestation and has identified surprising changes in extreme cold events in southern South America as well as in the deforested region. In particular, extreme cold events became both more frequent and colder in the western Amazon, and less frequent in the eastern Amazon. Deforestation of the Amazon also caused increases in the frequency and intensity of extreme cold events in southern South America. The southern La Plata Basin (roughly bounded by 15°–35°S, between the Andes and Brazilian Highlands) seems particularly vulnerable to changes in extreme cold events. This region is an important producer of soybean, rice, sunflower, wheat, and maize, and has itself experienced significant deforestation in the past few decades. Further research should investigate the combined effects of local deforestation and Amazon deforestation on extreme events in the La Plata Basin. We emphasize that the extreme cold events presented in this study are defined on the basis of the statistical distribution of minimum daily temperature and not on an absolute temperature threshold (e.g., 0°C). Thus, the consequences of extreme cold events in the Amazon, where temperatures are always well above freezing, will be very different from in the La Plata Basin, where T05 can be near the freezing mark (Fig. 3a).

Another important issue requiring further study is the mechanism by which deforestation causes changes in extreme events in southern South America. We have proposed a particular series of events whereby (i) deforestation causes the Amazon to warm, increasing the north–south temperature gradient; (ii) the westerly jet at roughly 20°S strengthens; (iii) a secondary circulation is set up, which reduces vertical velocities south of the jet and increases vertical velocities north of the jet; and (iv) this secondary circulation enhances low-level anticyclonic vorticity south of the jet and causes midlevel cooling north of the jet. Both of these factors act to strengthen the surface anticyclones associated with extreme cold events. Future work should investigate this chain of events with detailed case studies.

In the future, the Amazon may also experience reduced atmospheric moisture associated with deforestation (Fig. 7b). While the mechanisms investigated here did not directly implicate a drying of the Amazon with increases in extreme cold events, reduced levels of humidity could have important implications for temperature. For example, drier conditions can contribute to a greater diurnal surface temperature range because of less heat capacity in soil, the litter layer, and sometimes plants, and less latent heat release due to condensation at night. Changes in the latent heat flux may initiate complex cloud feedbacks, affecting both shortwave and longwave radiation (Medvigy et al. 2011). Assessing how these factors impact the global circulation and the intensity, frequency, and northward penetration of Southern Hemisphere baroclinic waves and cold fronts will require additional, carefully designed numerical experiments.

It is possible that Amazon deforestation is affecting extreme events in other parts of the world; however, such changes would be difficult to identify in our simulations because a coarser grid mesh (~200 km) was used outside of South America. This coarse mesh acts to smooth out extremes. In the future, it would be interesting to investigate the Midwest United States and east Africa in particular because Avissar and Werth (2005) have previously found that these regions’ hydroclimates were sensitive to deforestation of the Amazon.

In reality, Amazon deforestation will occur in concert with changes in greenhouse gas concentrations, changes in SSTs and sea ice, changes in solar forcing, and other changes. Our analysis has been simplified in the sense that we have modified land cover without changing any other boundary conditions or forcings. As computational resources increase, we will be able to explore the combined effects of deforestation and other changes.

Another important issue involves the timing and spatial pattern of Amazon deforestation. Over the next 50 yr, a total of 40% of the Amazon may be deforested (Soares-Filho et al. 2006). Our previous study (Medvigy et al. 2011) suggested that this level of deforestation would also impact the frequency of extreme cold events in southern South America, although not to the same degree as a total deforestation scenario. Additional work should be carried out to investigate how specific patterns of deforestation affect extreme events. Finally, we note that the methods discussed in this work can also be used to identify significant changes in extreme hot events or extreme precipitation events. A preliminary analysis suggests that such changes may occur in the deforested region but are weak elsewhere in South America. Additional work is necessary to determine if other regions of the globe may experience changes in extremes.

Acknowledgments

The authors gratefully acknowledge support from NSF Award 0902197. We also thank Marcos Longo and Gabriela Müller and three anonymous reviewers for their insightful comments. NCEP reanalysis 2 data were provided by the NOAA/OAR/ESRL PSD, Boulder, Colorado, from their Web site at http://www.esrl.noaa.gov/psd/. Some of the simulations presented in this article were performed on computational resources supported by the PICSciE-OIT High Performance Computing Center and Visualization Laboratory at Princeton University.

REFERENCES

REFERENCES
Avissar
,
R.
, and
D.
Werth
,
2005
:
Global hydroclimatological teleconnections resulting from tropical deforestation
.
J. Hydrometeor.
,
6
,
134
145
.
Barros
,
V. R.
,
A. M.
Grimm
, and
M. E.
Doyle
,
2002
:
Relationship between temperature and circulation in southeastern South America and its influence from El Niño and La Niña events
.
J. Meteor. Soc. Japan
,
80
,
21
32
.
Bell
,
J. L.
,
L. C.
Sloan
, and
M. A.
Snyder
,
2004
:
Regional changes in extreme climatic events: A future climate scenario
.
J. Climate
,
17
,
81
87
.
Caesar
,
J.
,
L.
Alexander
, and
R.
Vose
,
2006
:
Large-scale changes in observed daily maximum and minimum temperatures: Creation and analysis of a new gridded data set
.
J. Geophys. Res.
,
111
,
D05101
,
doi:10.1029/2005JD006280
.
Costa
,
M. H.
, and
J. A.
Foley
,
2000
:
Combined effects of deforestation and doubled atmospheric CO2 concentrations on the climate of Amazonia
.
J. Climate
,
13
,
18
34
.
Costa
,
M. H.
,
S. N. M.
Yanagi
,
P. J. O. P.
Souza
,
A.
Ribeiro
, and
E. J. P.
Rocha
,
2007
:
Climate change in Amazonia caused by soybean cropland expansion, as compared to caused by pastureland expansion
.
Geophys. Res. Lett.
,
34
,
L07706
,
doi:10.1029/2007GL029271
.
Diffenbaugh
,
N. S.
,
J. S.
Pal
,
R. J.
Trapp
, and
F.
Giorgi
,
2005
:
Fine-scale processes regulate the response of extreme events to global climate change
.
Proc. Natl. Acad. Sci. USA
,
102
,
15 774
15 778
.
Findell
,
K. L.
,
T. R.
Knutson
, and
P. C. D.
Milly
,
2006
:
Weak simulated extratropical responses to complete tropical deforestation
.
J. Climate
,
19
,
2835
2850
.
Gandu
,
A. W.
,
J. C. P.
Cohen
, and
J. R. S.
Souza
,
2004
:
Simulation of deforestation in eastern Amazonia using a high-resolution model
.
Theor. Appl. Climatol.
,
78
,
123
135
.
Garreaud
,
R. D.
,
1999
:
Cold air incursions over subtropical and tropical South America: A numerical case study
.
Mon. Wea. Rev.
,
127
,
2823
2853
.
Garreaud
,
R. D.
,
2000
:
Cold air incursions over subtropical South America: Mean structure and dynamics
.
Mon. Wea. Rev.
,
128
,
2544
2559
.
Gash
,
J. H. C.
, and
C. A.
Nobre
,
1997
:
Climatic effects of Amazonian deforestation: Some results from ABRACOS
.
Bull. Amer. Meteor. Soc.
,
78
,
823
830
.
Gedney
,
N.
, and
P. J.
Valdes
,
2000
:
The effect of Amazonian deforestation on the Northern Hemisphere circulation and climate
.
Geophys. Res. Lett.
,
27
,
3053
3056
.
Hahmann
,
A. N.
, and
R. E.
Dickinson
,
1997
:
RCCM2-BATS model over tropical South America: Applications to tropical deforestation
.
J. Climate
,
10
,
1944
1964
.
Hall
,
N. M. J.
,
B. J.
Hoskins
,
P. J.
Valdes
, and
C. A.
Senior
,
1994
:
Storm tracks in a high-resolution GCM with doubled carbon dioxide
.
Quart. J. Roy. Meteor. Soc.
,
120
,
1209
1230
.
Hasler
,
N.
,
D.
Werth
, and
R.
Avissar
,
2009
:
Effects of tropical deforestation on global hydroclimate: A multimodel ensemble analysis
.
J. Climate
,
22
,
1124
1141
.
Henderson-Sellers
,
A.
,
R. E.
Dickinson
,
T. B.
Durbidge
,
P. J.
Kennedy
,
K.
McGuffie
, and
A. J.
Pitman
,
1993
:
Tropical deforestation—Modeling local-scale to regional-scale climate change
.
J. Geophys. Res.
,
98
(
D4
),
7289
7315
.
Hollander
,
M.
, and
D. A.
Wolfe
,
1999
:
Nonparametric Statistical Methods. John Wiley and Sons, 787 pp
.
Kalnay
,
E.
, and
Coauthors
,
1996
:
The NCEP/NCAR 40-Year Reanalysis Project
.
Bull. Amer. Meteor. Soc.
,
77
,
437
471
.
Kummerow
,
C.
,
W.
Barnes
,
T.
Kozu
,
J.
Shiue
, and
J.
Simpson
,
1998
:
The Tropical Rainfall Measuring Mission (TRMM) sensor package
.
J. Atmos. Oceanic Technol.
,
15
,
809
817
.
Kunkel
,
K. E.
,
X.-Z.
Liang
, and
J.
Zhu
,
2010
:
Regional climate model projections and uncertainties of U.S. summer heat waves
.
J. Climate
,
23
,
4447
4458
.
Kushner
,
P. J.
,
I. M.
Held
, and
T. L.
Delworth
,
2001
:
Southern Hemisphere atmospheric circulation response to global warming
.
J. Climate
,
14
,
2238
2249
.
Kwiatkowski
,
D.
,
P. C. B.
Phillips
,
P.
Schmidt
, and
Y.
Shin
,
1992
:
Testing the null hypothesis of stationarity against the alternative of unit root
.
J. Econom.
,
54
,
159
178
.
Lau
,
K.-M.
, and
C.-P.
Chang
,
1987
:
Planetary scale aspects of the winter monsoon and atmospheric teleconnections. Monsoon Meteorology, C.-P. Change and T. N. Krishnamurti, Eds., Oxford University Press, 161–202
.
Lean
,
J.
, and
P. R.
Rowntree
,
1993
:
A GCM simulation of the impact of Amazonian deforestation on climate using an improved canopy representation
.
Quart. J. Roy. Meteor. Soc.
,
119
,
509
530
.
Ljung
,
G. M.
, and
G. E. P.
Box
,
1978
:
On a measure of lack of fit in time series models
.
Biometrika
,
65
,
297
303
.
Lloyd
,
I. D.
, and
G. A.
Vecchi
,
2010
:
Submonthly Indian Ocean cooling events and their interaction with large-scale conditions
.
J. Climate
,
23
,
700
716
.
Marengo
,
J.
,
A.
Cornejo
,
P.
Satyamurty
, and
C.
Nobre
,
1997a
:
Cold surges in tropical and extratropical South America: The strong event in June 1994
.
Mon. Wea. Rev.
,
125
,
2759
2786
.
Marengo
,
J.
,
C.
Nobre
, and
A.
Culf
,
1997b
:
Climatic impacts of “friagens” in forested and deforested regions in the Amazon Basin
.
J. Appl. Meteor.
,
36
,
1553
1566
.
Medvigy
,
D.
,
R. L.
Walko
, and
R.
Avissar
,
2008
:
Modeling interannual variability of the Amazon hydroclimate
.
Geophys. Res. Lett.
,
35
,
L15817
,
doi:10.1029/2008GL034941
.
Medvigy
,
D.
,
R. W.
Walko
,
M. J.
Otte
, and
R.
Avissar
,
2010
:
The Ocean–Land–Atmosphere Model (OLAM): Optimization and evaluation of simulated radiative fluxes and precipitation
.
Mon. Wea. Rev.
,
138
,
1923
1939
.
Medvigy
,
D.
,
R. L.
Walko
, and
R.
Avissar
,
2011
:
Effects of deforestation on spatiotemporal distributions of precipitation in South America
.
J. Climate
,
24
,
2147
2163
.
Meehl
,
G. A.
, and
Coauthors
,
2007
:
Global climate projections. Climate Change 2007: The Physical Science Basis, S. Solomon et al., Eds., Cambridge University Press, 747–845
.
Müller
,
G. V.
,
2010
:
Temperature decrease in the extratropics of South America in response to a tropical forcing during the austral winter
.
Ann. Geophys.
,
28
,
1
9
.
Müller
,
G. V.
, and
G. J.
Berri
,
2007
:
Atmospheric circulation associated with persistent generalized frosts in central-southern South America
.
Mon. Wea. Rev.
,
135
,
1268
1289
.
Müller
,
G. V.
,
T.
Ambrizzi
, and
M. N.
Núñez
,
2005
:
Mean atmospheric circulation leading to generalized frosts in central southern South America
.
Theor. Appl. Climatol.
,
82
,
95
112
.
Olson
,
J. S.
,
1994a
:
Global ecosystem framework-definitions. USGS EROS Data Center Internal Rep., 37 pp
.
Olson
,
J. S.
,
1994b
:
Global ecosystem framework-translation strategy. USGS EROS Data Center Internal Rep., 39 pp
.
Ramos da Silva
,
R.
,
D.
Werth
, and
R.
Avissar
,
2008
:
Regional impacts of future land cover changes on the Amazon basin wet-season climate
.
J. Climate
,
21
,
1153
1170
.
R Development Core Team
, cited
2008
:
R: A language and environment for statistical computing. R Foundation for Statistical Computing. [Available online at http://www.R-project.org.]
Reynolds
,
R. W.
,
N. A.
Rayner
,
T. M.
Smith
,
T. C.
Stokes
, and
W.
Wang
,
2002
:
An improved in situ and satellite SST analysis for climate
.
J. Climate
,
15
,
1609
1625
.
Rind
,
D.
,
M.
Chandler
,
J.
Lerner
,
D. G.
Martinson
, and
X.
Yuan
,
2001
:
Climate response to basin-specific changes in latitudinal temperature gradients and implications for sea ice variability
.
J. Geophys. Res.
,
106
(
D17
),
20 161
20 173
.
Royston
,
P.
,
1982
:
An extension of Shapiro and Wilk’s W test for normality to large samples
.
Appl. Stat.
,
31
,
115
124
.
Sampaio
,
G.
,
C.
Nobre
,
M. H.
Costa
,
P.
Satyamurty
,
B. S.
Soares-Filho
, and
M.
Cardoso
,
2007
:
Regional climate change over eastern Amazonia caused by pasture and soybean cropland expansion
.
Geophys. Res. Lett.
,
34
,
L17709
,
doi:10.1029/2007GL030612
.
Schultz
,
D. M.
,
W. E.
Bracken
, and
L. F.
Bosart
,
1998
:
Planetary- and synoptic-scale signatures associated with Central American cold surges
.
Mon. Wea. Rev.
,
126
,
5
27
.
Soares-Filho
,
B. S.
, and
Coauthors
,
2006
:
Modelling conservation in the Amazon basin
.
Nature
,
440
,
520
523
.
Uccellini
,
L. W.
, and
D. R.
Johnson
,
1979
:
The coupling of upper- and lower-tropospheric jet streaks and implications for the development of severe convective storms
.
Mon. Wea. Rev.
,
107
,
682
703
.
Walko
,
R. L.
, and
R.
Avissar
,
2008a
:
The Ocean–Land–Atmosphere Model (OLAM): Shallow water tests
.
Mon. Wea. Rev.
,
136
,
4033
4044
.
Walko
,
R. L.
, and
R.
Avissar
,
2008b
:
The Ocean–Land–Atmosphere Model (OLAM): Formulation and tests of the nonhydrostatic dynamic core
.
Mon. Wea. Rev.
,
136
,
4045
4062
.
Walko
,
R. L.
, and
R.
Avissar
,
2011
:
A direct method for constructing refined regions in unstructured conforming triangular–hexagonal computational grids: Application to OLAM
.
Mon. Wea. Rev.
,
139
,
3923
3937
.
Walko
,
R. L.
, and
Coauthors
,
2000
:
Coupled atmosphere–biophysics–hydrology models for environmental modeling
.
J. Appl. Meteor.
,
39
,
931
944
.
Werth
,
D.
, and
R.
Avissar
,
2002
:
The local and global effects of Amazon deforestation
.
J. Geophys. Res.
,
107
,
8087
,
doi:10.1029/2001JD000717
.
Yin
,
J. H.
,
2005
:
A consistent poleward shift of the storm tracks in simulations of 21st century climate
.
Geophys. Res. Lett.
,
32
,
L18701
,
doi:10.1029/2005GL023684
.